A PyTorch transplant of Mask R-CNN from the Keras era

What it does

The repo implements instance segmentation—drawing bounding boxes and pixel-level masks around every object in an image—using a ResNet101 backbone and Feature Pyramid Network. In practice, it is a PyTorch rewrite of Matterport’s widely used Keras/TensorFlow implementation, translating the same architecture and training logic into a different framework. The code exposes the full detection pipeline, from anchor filtering through bounding-box refinement to final mask composition.

The interesting bit

Because it targets PyTorch 0.3, the authors stitched in third-party CUDA kernels for Non-Maximum Suppression and RoIAlign by hand. That makes the repo as much a study in early PyTorch ecosystem plumbing as in computer vision.

Key highlights

• Generates both bounding boxes and segmentation masks per object instance.
• Trains and evaluates on MS COCO with pretrained ImageNet and COCO weights available.
• Runs inference on CPU or GPU.
• Detection scores (0.347 bbox AP, 0.296 segm AP) match the Keras version it is based on, though they sit below the original paper’s reported numbers.

Caveats

• Targets PyTorch 0.3, which is long obsolete; do not expect modern PyTorch compatibility without surgery.
• Requires manually compiling CUDA kernels for NMS and RoIAlign against specific GPU architectures.
• The “from scratch” training results are listed as “t.b.a.” in the README, so it is unclear how the model performs without Keras-transferred weights.

Verdict

Worth a look if you are maintaining legacy PyTorch 0.3 code or studying how early frameworks bridged the gap before standardized object-detection primitives arrived. Everyone else should probably use modern torchvision detection models and move on.